Hugging Face's logo

Datasets:
Fanoni
/
medical-ehr-imaging

Modalities:
Image
Formats:
imagefolder
Size:
1K - 10K
Libraries:
Datasets
Dataset card Data Studio Files
xet
Community
medical-ehr-imaging / validation_report.md
dafe smith
Upload folder using huggingface_hub
9c82964 verified
preview code
|
raw
history blame contribute delete
8.65 kB

Medical Imaging Dataset Validation Report

Date: November 23, 2025 Validation Tool Version: 1.0 Base Directory: /Users/dafesmith/Documents/repo/NeMo-agent/medical_ehr/data/imaging/

Executive Summary

Metric Value
Total Datasets 4
Total Files 633
Valid Files 620
Invalid Files 13
Total Size 385.38 MB
Validation Rate 97.95%

Datasets Downloaded

1. Chest X-ray (Pneumonia Detection)

Property Value
Source HuggingFace: hf-vision/chest-xray-pneumonia
Location chest_xray/pneumonia/
Total Files 203
Valid Files 203 (100%)
Total Size 313.03 MB
Image Format PNG
Image Resolution Variable (avg ~1500x1200 pixels)

File Types:

  • PNG images: 200
  • JSON metadata: 2
  • CSV labels: 1

Label Distribution:

  • Label 0 (Normal): 200 images

Sample Files:

  • 0_00124.png (1430x1128, 1.2 MB)
  • 0_00130.png (1654x1368, 1.8 MB)
  • 0_00118.png (1828x1511, 1.9 MB)

2. Brain MRI (Alzheimer Detection)

Property Value
Source HuggingFace: Falah/Alzheimer_MRI
Location brain_mri/alzheimer/
Total Files 203
Valid Files 203 (100%)
Total Size 2.97 MB
Image Format PNG
Image Resolution 128x128 pixels

File Types:

  • PNG images: 200
  • JSON metadata: 2
  • CSV labels: 1

Label Distribution:

  • Label 0 (Mild Demented): 29 images
  • Label 1 (Moderate Demented): 2 images
  • Label 2 (Non Demented): 102 images
  • Label 3 (Very Mild Demented): 67 images

Sample Files:

  • 2_00007.png (128x128, 14.7 KB)
  • 0_00118.png (128x128, 14.7 KB)
  • 3_00070.png (128x128, 16.6 KB)

3. Dermatology (Skin Cancer Classification)

Property Value
Source HuggingFace: marmal88/skin_cancer
Location dermatology/skin_cancer/
Total Files 203
Valid Files 203 (100%)
Total Size 68.77 MB
Image Format PNG
Image Resolution 600x450 pixels

File Types:

  • PNG images: 200
  • JSON metadata: 2
  • CSV labels: 1

Label Distribution:

  • Actinic Keratoses: 200 images

Additional Metadata Fields:

  • image_id, lesion_id, dx_type, age, sex, localization

Sample Files:

  • actinic_keratoses_00056.png (600x450, 310 KB)
  • actinic_keratoses_00042.png (600x450, 392 KB)
  • actinic_keratoses_00095.png (600x450, 383 KB)

4. DICOM Samples (Multi-modality Test Files)

Property Value
Source PyDICOM GitHub Repository
Location dicom_samples/
Total Files 24
Valid Files 11 (45.8%)
Invalid Files 13
Total Size 0.60 MB
Format DICOM (.dcm)

File Types:

  • DICOM files: 23
  • JSON metadata: 1

Successfully Validated DICOM Files:

Filename Modality Study Date Size
MR_small.dcm MR 20040826 9.6 KB
MR_small_bigendian.dcm MR 20040826 9.7 KB
MR_small_implicit.dcm MR 20040826 9.7 KB
MR_small_RLE.dcm MR 20040826 7.8 KB
MR_small_expb.dcm MR 20040826 9.8 KB
CT_small.dcm CT - 39.7 KB
ExplVR_BigEnd.dcm - - 15.4 KB
VR-2022.dcm - - 258 KB
J2K_pixelrep_mismatch.dcm - - 138.5 KB

Files with Validation Issues (codec/format issues, not corruption):

  • JPEG Extended transfer syntax (requires specific codecs)
  • JPEG-LS files (requires pyjpegls or GDCM)
  • Files without DICM header (legacy format)
  • JPEG2000 with bit depth issues

Directory Structure 

/Users/dafesmith/Documents/repo/NeMo-agent/medical_ehr/data/imaging/
├── brain_mri/
│   └── alzheimer/
│       ├── images/          (200 PNG files)
│       └── metadata/        (labels.csv, labels.json, summary.json)
├── chest_xray/
│   ├── images/              (empty - not used)
│   ├── metadata/            (empty - not used)
│   └── pneumonia/
│       ├── images/          (200 PNG files)
│       └── metadata/        (labels.csv, labels.json, summary.json)
├── ct_scans/
│   └── lidc_idri/           (empty - requires TCIA download)
├── dermatology/
│   └── skin_cancer/
│       ├── images/          (200 PNG files)
│       └── metadata/        (labels.csv, labels.json, summary.json)
├── dicom_samples/
│   ├── brainix/             (empty - requires premium access)
│   ├── manix/               (empty - requires premium access)
│   └── *.dcm                (23 DICOM test files)
├── medical_docs/
│   └── ocr_samples/         (empty - future use)
├── download_medical_imaging.py
├── download_chest_xray.py
├── validate_datasets.py
├── download_summary.json
├── validation_results.json
└── validation_report.md     (this file)

Issues Encountered

1. HuggingFace Dataset Script Deprecation

Issue: The original alkzar90/NIH-Chest-X-ray-dataset uses deprecated dataset scripts.

Error: RuntimeError: Dataset scripts are no longer supported

Resolution: Used alternative dataset hf-vision/chest-xray-pneumonia with modern Parquet format.

2. OsiriX DICOM Library Access

Issue: OsiriX sample datasets (BRAINIX, MANIX) require premium membership.

Resolution: Downloaded alternative DICOM test files from PyDICOM GitHub repository.

3. DICOM Codec Dependencies

Issue: Some DICOM files require specialized codecs (JPEG-LS, GDCM) for pixel data extraction.

Error Types:

  • "JPEG Extended only supported by Pillow if Bits Allocated = 8"
  • "Missing required dependencies: GDCM, pyjpegls"

Resolution: Files are valid DICOM, but pixel data extraction requires additional libraries. For testing purposes, the available files with standard transfer syntaxes are sufficient.

4. Kaggle CLI Not Available

Issue: Kaggle CLI not installed, preventing direct NIH Chest X-ray dataset download.

Resolution: Used HuggingFace as primary source.

Validation Details

Image Validation Checks

  1. File Integrity: PIL Image.verify() for PNG/JPG files
  2. Format Verification: Confirmed PNG format and RGB mode
  3. Size Measurement: File size and image dimensions
  4. Metadata Presence: Checked for labels.csv and labels.json

DICOM Validation Checks

  1. DICM Magic Number: Checked for DICOM preamble
  2. Metadata Extraction: Patient ID (anonymized), Modality, Study Date
  3. Pixel Data: Attempted pixel array extraction where possible

Recommendations

Immediate Next Steps

  1. Install DICOM Codecs (optional):

    pip install pylibjpeg pylibjpeg-libjpeg pyjpegls

  2. Download More Chest X-rays: Current sample has limited label diversity (all label 0). Consider downloading balanced dataset.

  3. Add CT Scan Data: LIDC-IDRI requires TCIA Data Retriever. Alternative: use Stanford AIMI datasets.

For Production Use

  1. PhysioNet Credentialing: Register for MIMIC-CXR access (224,316 chest X-rays)
  2. TCIA Registration: Access LIDC-IDRI (1,018 CT scans with annotations)
  3. VinDr-CXR: Consider Vietnamese chest X-ray dataset (18,000 images)

Data Augmentation

Consider augmenting the current datasets:

  • Random rotations, flips, brightness adjustments for training
  • Resize consistency (all images to standard size like 224x224)

Files Created

File Purpose Size
download_medical_imaging.py Main download script 6 KB
download_chest_xray.py ChestX-ray14 download (legacy) 3 KB
validate_datasets.py Validation script 6 KB
download_summary.json Download metadata 1 KB
validation_results.json Detailed validation output 11 KB
validation_report.md This report 8 KB

Appendix: Dataset Sources

Dataset Source License Registration Required
Chest X-ray Pneumonia HuggingFace CC0 No
Alzheimer MRI HuggingFace Unknown No
Skin Cancer HuggingFace CC BY-NC-SA 4.0 No
DICOM Samples PyDICOM MIT No
LIDC-IDRI TCIA CC BY 3.0 Yes (TCIA)
MIMIC-CXR PhysioNet PhysioNet Yes (PhysioNet)
OsiriX Samples OsiriX Premium Yes (Paid)

Report Generated: 2025-11-23T13:37:06 Validation Script: validate_datasets.py Download Script: download_medical_imaging.py